GRAPE: Gossip-Based Representation Alignment Using Prototypical Embeddings for Heterogeneous IoT Device
| Tipo de publicação: | Inproceedings |
| Citação: | IEEE WCNC |
| Ano: | 2026 |
| Mês: | April |
| Publisher: | IEEE |
| Location: | Kuala Lumpur |
| Resumo: | Gossip learning (GL) introduces a fully decentralized framework for training machine learning models across large-scale networks of heterogeneous IoT devices, leveraging peer-to-peer communication to enable collaborative learning without centralized coordination or raw data exchange. However, in real-world IoT environments, devices commonly operate under non-IID data distributions and often contains different model architectures, computational power, and storage capacity which leads to model heterogeneity. In such settings, conventional GL fails to converge reliably, as weight averaging is incompatible with heterogeneous architectures, while existing alternatives still rely on centralized coordination. To address these challenges, we introduce GRAPE, a fully decentralized, prototype-based GL framework that tackles both data and model heterogeneity. Instead of exchanging full model parameters, participating devices communicate compact class prototypes that serve as semantic summaries of local knowledge. We propose a contrastive prototype alignment objective that enables devices to refine their learned representations by attracting semantically consistent prototypes while repelling inconsistent ones during peer-to-peer exchanges. This approach enforces consistent representations across heterogeneous models within a unified embedding space. Extensive experiments show that GRAPE outperforms standard GL baselines, achieving higher accuracy, improved robustness to heterogeneity, and lower communication costs. |
| Palavras-chave: | Distributed Learning, Gossip Learning, Internet of Things (IoT) |
| Autores | |
| Adicionado por: | [] |
| Total mark: | 0 |
|
Anexos
|
|
|
|
|
Notas
|
|
|
|
|
|
Tópicos
|
|
|
|
|
